Data processing method and system for identifying false recruitment

Abstract

The present invention relates to a data processing method and system for identifying fraudulent recruitment. The method comprises: extracting target recruiting company information, target position information, and interview information from recruitment interview information provided by a user; crawling information from the network based on multiple indicator reference contents at the company level, position level, and interview level based on the extracted target recruiting company information, target position information, and interview information; processing the crawled information at each level to obtain multiple indicator data at corresponding levels; and normalizing each indicator data into feature data of a machine learning model, and combining the feature data corresponding to the indicator data at each level to form prediction samples for the machine learning models at each level, wherein the machine learning models at each level are used to assess the corresponding level risk of fraudulent recruitment. The method and system provided by the present invention improve the accuracy of identifying fraudulent recruitment.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data processing method and system for identifying false recruitment. Background Art

[0002] Fake recruitment generally refers to fraudulent recruitment activities where job postings online or in talent markets do not reflect actual job openings. This can include both fake recruitment by legitimate companies and illegal recruitment by fraudulent companies. Some legitimate companies, for purposes such as collecting talent information and promoting their reputation, create fictitious positions and post recruitment information online and / or in talent markets. This fraudulent recruitment not only delays job seekers' opportunities and compromises personal information security, but also consumes and wastes time on public job search platforms, undermining their credibility. Illegal recruitment by fraudulent companies can result in property and personal harm to job seekers, such as through excessive fees, pyramid schemes, and involvement in pornography and drug trafficking. It can also harm the interests of third parties, such as the impersonating legitimate companies and the recruitment platforms that post job postings for them.

[0003] In order to identify these false recruitment information, the industry has made a lot of efforts. The Chinese patent application with publication number CN113704409 A and invention name "A method for detecting false recruitment information based on cascade forest" provides a false recruitment information detection method, which is based on the cascade forest algorithm of decision trees and uses the job data released by online recruitment platforms to establish a model for false recruitment prediction. The Chinese patent application with publication number CN 113506084 A and invention name "A method for detecting false recruitment positions based on deep learning" provides a false recruitment position detection method, which collects false recruitment information from online recruitment platforms or recruitment apps, processes the false recruitment information into sample data for training models, obtains a detection model through training, and uses the detection model to detect false recruitment information on the online recruitment platforms or recruitment apps.

[0004] The technical solutions presented above demonstrate that, currently, when identifying fake job postings, the training data used to train the recognition model comes from fake job postings and / or fake recruitment information collected from online recruitment platforms. This data source is narrow, and the identification dimension is single. As the aforementioned description of fake job postings indicates, fraudulent job postings are multifaceted. If identification is limited to the job title, a large number of fake job postings will inevitably be missed. Therefore, a solution is urgently needed to identify fake job postings from multiple perspectives and dimensions. Summary of the Invention

[0005] In response to the technical problems existing in the prior art, the present invention proposes a data processing method and system for identifying false recruitment, which obtains information used to identify false recruitment from multiple aspects and dimensions, thereby improving the accuracy of false recruitment identification.

[0006] In order to solve the above technical problem, according to one aspect of the present invention, the present invention provides a data processing method for identifying false recruitment, which includes the following steps:

[0007] Extract target recruiting company information, target position information and interview information from the recruitment interview information provided by the user;

[0008] Based on the extracted target recruitment company information, target position information and interview information, information is crawled on the network according to multiple indicator reference contents at the company level, position level and interview level;

[0009] Based on the crawled information at each level, multiple indicator data of the corresponding level are obtained; and

[0010] The indicator data is standardized into feature data of the machine learning model, and the feature data corresponding to the indicator data at each level are combined to form prediction samples of the machine learning models at each level, wherein the machine learning models at each level are used to assess the corresponding level of risk of false recruitment.

[0011] According to one aspect of the present invention, the present invention provides a data processing system for identifying false recruitment, wherein the system includes a user information acquisition module, a data collection module, an indicator data generation module and a prediction sample generation module, wherein the user information acquisition module is configured to receive recruitment interview information provided by the user, and extract target recruitment enterprise information, target position information and interview information from the user recruitment interview information; the data collection module is connected to the Internet and connected to the user information acquisition module, and is configured to crawl information in the network according to multiple indicator reference contents at the enterprise level, position level and interview level based on the extracted target recruitment enterprise information, target position information and interview information; the indicator data generation module is connected to the data collection module, and obtains multiple indicator data of corresponding levels based on the crawled information at each level; the prediction sample generation module is connected to the indicator data generation module, and is configured to normalize the indicator data into feature data of the machine learning model, and combine the feature data corresponding to the indicator data at each level to form prediction samples of the machine learning model at each level, wherein the machine learning models at each level are used to assess the corresponding level risk of false recruitment.

[0012] It can be seen from the above methods and systems provided by the present invention that the present invention makes up for the lack of predecessors' failure to actively search for information and identify risks: traditional algorithms only use information related to the job description to determine whether it is a false recruitment, while the present invention actively searches for other relevant information, such as company information, related company information, target position information, related position information, etc., which increases the accuracy of the algorithm's judgment. The present invention makes up for the lack of predecessors' failure to predict risks in the interview notification stage: existing false recruitment detection schemes are earlier than the interview notification stage, and fail to effectively detect false recruitments that only reveal risks in the interview notification stage. The present invention analyzes and evaluates the interview information provided by the job seeker after the job seeker receives the interview notification from the target company. This timing is both the last line of defense for the job seeker before going to the interview and the best time to obtain the most sufficient data to predict risks, thereby helping the job seeker to avoid risks in a targeted manner to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Below, the preferred embodiments of the present invention will be further described in detail with reference to the accompanying drawings, in which:

[0014] Figure 1 is a flow chart of a method for identifying fake recruitment according to an embodiment of the present invention;

[0015] Figure 2 is a flow chart of risk assessment according to one embodiment of the present invention;

[0016] Figure 3 is a flow chart of risk assessment according to another embodiment of the present invention;

[0017] Figure 4 is a flow chart of risk assessment according to yet another embodiment of the present invention;

[0018] Figure 5 is a flow chart of risk assessment according to another embodiment of the present invention;

[0019] Figure 6 is a flow chart of a method for processing test sample data according to one embodiment of the present invention;

[0020] Figure 7 is a flow chart of a false recruitment warning method according to an embodiment of the present invention;

[0021] Figure 8 This is a functional block diagram of a fake recruitment identification system according to an embodiment of the present invention;

[0022] Figure 9 This is a partial principle block diagram of a fake recruitment identification system provided according to one embodiment of the present invention;

[0023] Figure 10This is a partial principle block diagram of a fake recruitment identification system provided according to an embodiment of the present invention;

[0024] Figure 11 is a functional block diagram of a user information acquisition module according to an embodiment of the present invention;

[0025] Figure 12 is a principle block diagram of a data processing system according to an embodiment of the present invention;

[0026] Figure 13 is a principle block diagram of a data collection module according to one embodiment of the present invention;

[0027] Figure 14 is a principle block diagram of an indicator data generation module according to an embodiment of the present invention;

[0028] Figure 15 is a principle block diagram of a model training module according to one embodiment of the present invention;

[0029] Figure 16 This is a functional block diagram of a false recruitment warning system according to an embodiment of the present invention;

[0030] Figure 17 is a principle block diagram of an early warning module according to an embodiment of the present invention;

[0031] Figure 18 is a principle block diagram of an early warning module according to another embodiment of the present invention; and

[0032] Figure 19 It is a principle block diagram of an early warning module according to another embodiment of the present invention. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0034] In the detailed description that follows, reference may be made to the various drawings that form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Each specific embodiment of the present application is described below in sufficient detail to enable a person of ordinary skill in the art to implement the technical solutions of the present application. It should be understood that other embodiments may be utilized or that structural, logical, or electrical changes may be made to the embodiments of the present application.

[0035] Figure 1 : is a flow chart of a method for identifying fake recruitment according to an embodiment of the present invention. In this embodiment, the method includes the following steps:

[0036] Step S1a, determines whether the recruitment interview information provided by the user is received. If the recruitment interview information provided by the user is received, step S2a is executed. If not, this step is repeated. In one embodiment, when the method is applied to an online recruitment platform, such as a recruitment website, a recruitment APP, etc., the online recruitment platform opens an interface to the user to facilitate the user to input the recruitment interview information through the interface. The interface can be implemented as a user interface, for example, and the user inputs the recruitment interview information through the user interface. In some embodiments, the user interface is also provided with multiple information segments, which are used to input the target recruitment company information, target position information and interview information respectively. The target recruitment company information includes, for example, information such as the company name and address. The target position information includes, for example, information such as the position name and work location. The interview information includes, for example, the interview location, interview time, interview notification method, interview format, such as oral test, written test and other information.

[0037] Step S2a extracts target company information, target position information, and interview information from the recruitment interview information provided by the user. In one embodiment, when the user enters various information via the information segment in the user interface, the information in the field is read from the information segment and processed to obtain one or more keywords. This processing may include, for example, removing stop words and performing word segmentation.

[0038] Step S3a, based on the extracted target recruiting company information, target position information and interview information, information is crawled on the network according to multiple indicators of company level, position level and interview level.

[0039] In one embodiment, enterprise-level indicators include static and dynamic indicators. Static indicators include, but are not limited to, one or more of the following: the number of employees, number of branches, registered capital, financing information, company history, annual turnover, business scope, and company type. Dynamic indicators include, but are not limited to, one or more of the following: legal proceedings, major investments received, social news events, and social media reviews of the target company. Position-level indicators include, but are not limited to, one or more of the following: job title, work location, work department, monthly salary range, job description, benefits, job type, experience requirements, educational requirements, school type, industry risk, and the number of repeated recruitments for the same position. Interview-level indicators include, but are not limited to, one or more of the following: interview location, interview time, interview notification method, number of interviews conducted, whether a written test was required, and the interview format. Based on these indicators, corresponding information is retrieved from the internet. For example, based on the name and address of the target company provided by the user, a web search using the name and / or address as keywords can retrieve the target company's official website, social media public account, public news reports, and so on.

[0040] In one embodiment, the target recruiting company's official website is searched and retrieved for company headcount information, information including "xx branch," business scope description, company category, financing information, company history description, annual turnover description, and registered capital information. If any of the above information is not retrieved from the target recruiting company's official website, the search continues online. For example, if the target recruiting company's official website does not contain information such as registered capital or company type, the target recruiting company's registration information can be retrieved from the industrial and commercial information network, and the registered capital and company type can be obtained from this information. For another example, if the target recruiting company's official website does not contain financing information, financing-related information can be obtained by querying the target recruiting company's social media accounts or public news reports. Comments, litigation information, and news about investment received by the target recruiting company over the past five years are identified on the web. To obtain position level indicator data, multiple recruitment platforms are searched for information related to the same and other positions posted by the target recruiting company. Information related to the position is also searched for information related to the same and other positions posted by companies of the same type as the target recruiting company. Information related to the position is retrieved from the target recruiting company's official website. Obtain job information posted by target recruiting companies within a certain period of time from the Internet.

[0041] Step S4a: Process the crawled information into indicator data of the corresponding level. The crawled information is processed according to the processing strategy of the specific indicator, and the corresponding information is standardized into specific indicator data according to various indicators. The indicator data can be a floating point number, a string vector or some encoding.

[0042] For example, processing of enterprise-level static metrics includes: Regarding the number of employees, the crawled target recruitment enterprise headcount data is converted into floating-point values. Regarding the number of branches, the number of branches is calculated based on the words "**branch" identified from the company's official website description and converted into a floating-point value. Regarding registered capital, the values ​​of the words "registered capital**" identified from the company's registration information are converted into floating-point values. Regarding business scope, keywords in the business scope description are identified and converted into a string vector. Regarding enterprise type, enterprise category information from the official website is converted into a string vector and matched against "state-owned enterprise," "foreign-invested enterprise," "joint venture," and "privately-owned enterprise." These options are converted into 3, 2, 1, and 0, respectively, and recorded. Regarding financing information, the net worth obtained from the enterprise's latest financial report is converted into a floating-point value. Regarding the number of years of operation, the company's founding date is subtracted from the current year to obtain the number of years of operation, which is then converted into a floating-point value. Regarding annual turnover, the annual turnover obtained from the enterprise's latest financial report is converted into a floating-point value.

[0043] The processing of dynamic indicators at the enterprise level includes: regarding corporate legal litigation events, first extract the litigation information of the target recruiting company in the most recent period (such as five years) from the acquired information; then use keywords such as "labor" and "defendant ** (name of the target recruiting company)" to perform regular matching on each piece of information. If the match is successful, it is recorded as 1, and if the match fails, it is recorded as 0; when the match is successful, the litigation time information is obtained from the litigation information and encoded. For example, when the litigation time is less than 1 year, it is encoded as 1; when the litigation time is less than 2 years and greater than 1 year, it is encoded as 0.8; when the litigation time is less than 3 years and greater than 2 years, it is encoded as 0.6; when the litigation time is less than 4 years and greater than 3 years, it is encoded as 0.5; when the litigation time is less than 5 years and greater than 4 years, it is encoded as 0.3; then multiply all the successfully matched codes by the litigation time code to obtain the litigation code; finally, all the litigation codes are added up to obtain the litigation event score of the enterprise.

[0044] Regarding the major investments received by the enterprise, first extract the news information related to the investment received in the most recent period (such as five years) from the captured information; then process the amount of investment received into floating-point values; then add up all the floating-point values ​​to obtain the total amount of major investments received by the target recruiting enterprise in the most recent period (such as five years); if there is none, the default value is 0.

[0045] Regarding social media company evaluations, we first extract multiple comments from the most recent period (e.g., five years) from the captured information; then process each comment according to the following process: remove HTML symbols from the comments through parsing methods; remove punctuation through regular expressions; use a stop word library to filter and remove all stop words; label the data text after the above cleaning; use a sentiment analysis model to analyze the text and output each comment as a comment text sentiment code of -1 (negative) or 1 (positive); then obtain the comment time information, and code it as 1 when the comment time is less than 1 year from the current time, code it as 0.8 when the comment time is less than 2 years and greater than 1 year from the current time, code it as 0.6 when the comment time is less than 3 years and greater than 2 years from the current time; When the comment time is less than 4 years and more than 3 years from the present, it is coded as 0.5; when the comment time is less than 5 years and more than 4 years from the present, it is coded as 0.3; then the sentiment code of all comment texts is multiplied by the time code to obtain the sentiment score of each comment. For example, a company has received a negative comment, and the time from the present is less than 4 years and more than 3 years, then the sentiment score of this comment is -1*0.5=-0.5; the company also received a positive comment, and the time from the present is less than 3 years and more than 2 years, then the sentiment score of this comment is 1*0.6=0.6; then the sentiment scores of all comments on the company are added together to obtain the company's social media evaluation score. For example, the social media evaluation score of the company in the above example is -0.5+0.6=0.1.

[0046] Position-level indicator data is processed as follows: For job titles, they are converted into string vectors. For work locations, the extracted work location information is first converted into a string vector. The corresponding geographic location is found using a map search engine. The owner / tenant names of the building at the location are then obtained and matched with the company name. If a match is found with the target recruiting company name, the code is coded as 0; otherwise, it is coded as 1. For work departments, the extracted work department information is first converted into a string vector. This is then matched with the company's official website. If a match is found on the official website, the code is coded as 0; otherwise, it is coded as 1. For monthly salary ranges, they are converted into floating-point values. For job descriptions, HTML symbols, spaces, and punctuation are first removed using parsing methods and regular expressions. Stop words are then identified and deleted using a stop word library. Finally, the remaining text is converted into a string vector. For benefits, HTML symbols, spaces, and punctuation are first removed using parsing methods and regular expressions. Stop words are then identified and deleted using a stop word library. Finally, the remaining text is converted into a string vector. Regarding job types, the crawled information is searched to see if there are characters such as "full-time," "part-time," "labor dispatch," and "hourly wage." If "full-time" is present, it is recorded as 3, "part-time" as 2, "labor dispatch" as 1, and "hourly wage" as 0. If none of the four are present, it is recorded as a null value. Regarding experience requirements, the crawled position-related information is first identified to see if it contains the string "experience." If so, the value before the "year" string is identified to obtain the required number of years of work and processed as a floating-point value. The minimum value is recorded. If not, the "fresh graduate" string is identified. If so, it is recorded as 0. Otherwise, it is recorded as a null value. Regarding educational requirements, the crawled position-related information is first identified to see if there is a description of educational requirements and matched against the strings "Master's degree or above," "Undergraduate degree," "College degree," "Technical secondary school or below," and "No requirements." The above options are processed as 4, 3, 2, 1, and 0, respectively, and recorded. For school type, we identify descriptions of the graduate school from the crawled job-related information and match them against strings such as "985" and "211". If "985" is identified, it is recorded as 2; if "211" is identified, it is recorded as 1; if both "985" and "211" are identified, it is also recorded as 1; if no matching fields are identified, it is recorded as 0. For job industry risk, we first match the crawled job description information with the "Industry Classification of the People's Republic of China" to obtain the industry to which it belongs; then, we search for related news using the industry as a keyword. If keywords such as "layoffs," "transformation," "market value evaporation," and "store closures" are found, it is defined as high risk and recorded as 1; otherwise, it is defined as low risk and recorded as 0.Regarding the number of repeated recruitments for the same position, we first crawl all the job positions posted by the target recruiting companies in the recent period (such as three years); then identify the duration of the same position posted online and convert it into a floating-point value on an annual basis; if a position is taken offline and then put back online, add up the online duration and convert it into a floating-point value.

[0047] The aforementioned target position's job title, location, department, monthly salary range, job description, benefits, job type, experience requirements, educational requirements, school type, industry risk, and number of repeated recruitments for the same position serve as depth indicators for a target position. A standardized collection of these indicators can serve as a prediction sample set for the target position. In another embodiment, while obtaining target position depth indicator data, the process also includes establishing an enterprise network recruitment information matrix based on position-level information, and obtaining, from this enterprise network recruitment information matrix, indicators related to the breadth of recruitment platforms and target positions, hereinafter referred to as breadth indicators.

[0048] The enterprise network recruitment information matrix includes one or more job information posted by the target enterprise and the qualification information of the recruitment platform that posted the job information; one or more job information posted by similar enterprises to the target enterprise and the qualification information of the recruitment platform that posted the job information. The jobs posted by the target enterprise include the target job and a second job different from the target job; and the recruitment platform includes one or more target recruitment platforms that post the target job provided by the target enterprise and one or more second recruitment platforms that post the second job information different from the target job.

[0049] The steps of obtaining the corresponding breadth index data based on the enterprise network recruitment information matrix include:

[0050] Calculate the qualification level vector of each recruitment platform based on the recruitment platform qualification information;

[0051] Count the number of positions posted on each recruitment platform from the enterprise network recruitment information matrix;

[0052] The engagement level of each recruitment platform is determined based on the number of positions; for example, the number of positions is used as the engagement level, or the number of positions is normalized and the normalized value is used as the engagement level.

[0053] The input level of each recruitment platform is used as the qualification level vector weight of the recruitment platform, and the qualification level weighted vector value of each recruitment platform is calculated. The qualification level weighted vector value of the recruitment platform is a breadth indicator.

[0054] The step of obtaining corresponding breadth index data based on the enterprise network recruitment information matrix also includes:

[0055] Searching for multiple second target positions that are the same as the target position and are provided by similar companies;

[0056] Calculate a position vector of the target position and each second target position based on the position information;

[0057] performing a clustering operation on the plurality of second target positions to obtain a second target position with the largest clustering result;

[0058] Calculate the first vector difference between the target position's position vector and the position vector of the second target position with the largest clustering result, and use the first vector difference as the target position's external consistency coefficient. The target position's external consistency coefficient is another breadth indicator.

[0059] The step of obtaining corresponding breadth index data based on the enterprise network recruitment information matrix also includes:

[0060] Calculate the second vector difference between the target position's position vector and the position vector of each second position provided by the target enterprise, and calculate the average of the second vector differences. Use the average of the second vector differences as the internal consistency coefficient of the target position. The internal consistency coefficient of the target position is another breadth indicator.

[0061] The step of obtaining corresponding breadth index data based on the enterprise network recruitment information matrix also includes:

[0062] Based on the average of the first vector difference and the second vector difference, the weights of the first vector difference and the second vector difference are determined according to their closeness to the target position vector, and the weighted average of the two is calculated as the consistency coefficient of the target position. The consistency coefficient of the target position is another breadth indicator.

[0063] In one embodiment, the depth index data and breadth index data obtained above are combined together as a prediction sample set for the target position.

[0064] The processing of interview-level indicator data includes: Regarding the interview location, first process the interview location information into a string vector; then find the corresponding geographic location information through a map search engine; then obtain the owner / tenant name of the building at the location and match it with the target recruiting company name; if the target recruiting company name is matched, it is coded as 0, otherwise it is 1. For the interview time, if the interview time information is between 8:00 and 18:00, it is recorded as 0, otherwise it is recorded as 1. For the interview notification method, if it is a phone call, it is coded as 1, if it is a text message, it is coded as 2, if it is an email, it is coded as 3, if it is a recruitment application (recruitment app), it is coded as 4, if it is a normal social software method, it is coded as 5, and multiple contact methods are allowed. Regarding whether there is a written test, if there is a written test, it is coded as 0, and if not, it is coded as 1.

[0065] In step S5a, the target enterprise is evaluated based on multiple indicators at various levels to determine the enterprise-level risk, position-level risk, and interview-level risk of fraudulent recruitment. In this embodiment, each risk level is assessed using a trained model. The model can employ algorithms such as decision trees, naive Bayes, multi-layer algorithms, k-nearest neighbor algorithms, random forest algorithms, and neural networks.

[0066] In this embodiment, the machine learning models at each level are trained using labeled samples at each level. The training of an enterprise-level machine learning model is used as an example to illustrate the process of training a machine learning model.

[0067] The training data for enterprise-level machine learning models includes a certain amount of enterprise sample data. Each enterprise sample data piece includes multiple indicator data labeled as either a fake recruiting enterprise or a real recruiting enterprise, with each indicator data piece serving as a feature data of the enterprise sample. For example, factors such as "enterprise headcount," "number of branches," "registered capital," "business scope," "enterprise type," "financing information," "enterprise age," "annual turnover," "enterprise legal proceedings," "significant investments received," and "social media company reviews" are used as independent variables (i.e., one-dimensional features of a sample). Each sample is manually labeled to determine whether it exhibits enterprise-level risk, such as labeling risky samples as "negative" or "positive" and non-risky samples as "positive" or "negative," thereby obtaining a sample set. This sample set should meet certain requirements of the training model, such as a balanced number of positive and negative samples and the same feature dimensions for each sample. 80% of the samples in the sample set are used as training samples, while 20% are used as validation samples, thus forming the training and validation sets, respectively.

[0068] Then, a model is constructed using any of the aforementioned decision tree, naive Bayesian, multi-layer, k-nearest neighbor, random forest, neural network, and other algorithms. The model performs supervised learning based on the training set samples according to the model algorithm to determine the probability that each sample contains enterprise-level risk.

[0069] Since this embodiment only sets two outcome types, "risky" and "no risk", the model uses a binary classification output. Assuming that the number of sample independent variables (i.e., the dimension of sample features) is n, the mathematical representation of the model is:

[0070] X={x1,x2,…,x n}

[0071] Y={y0,y1}

[0072] The X is the feature set of a sample input to the model, and each x i is a one-dimensional feature. For example: x1 is the number of employees in the enterprise, x2 is the number of branches, x3 is the registered capital, ..., x n is the annual turnover. Y is the classification set of the model, y i ∈0,1, in this embodiment, y0 represents "no risk" and y1 represents "risky". The model is based on a given sample X i , calculated as label c j The probability p(c j |X i ), the output of the model is:

[0073]

[0074] in Represents the classification prediction value, p(c j |X i ) represents the probability of each classification under a given sample. In one embodiment, a comparison threshold is set, such as 0.6-0.9. In this embodiment, a threshold of 0.8 is used for illustration. When judging whether there is an enterprise-level risk, if the predicted probability of y1 output by the algorithm model exceeds 0.8, it can be confirmed that the enterprise corresponding to the sample has an enterprise-level risk. Among them, the threshold can be an empirical value obtained by repeatedly calculating the model based on practical data, or it can be further adjusted as the model is iterated. The trained model is verified using the validation set sample. When the model meets the evaluation criteria, it can be used for online risk assessment.

[0075] The training process of the position-level machine learning model and the interview-level machine learning model is similar to the above process and will not be repeated here.

[0076] In step S4a, to ensure that the normalized data meets the model input requirements, i.e., a sample consists of multi-dimensional features with the same number of features, the indicator data at each level is standardized into floating-point numbers, string vectors, or encodings. These data are then combined into the model's prediction samples according to the input requirements of each level of model. Then, in step S5a, the prediction samples at each level are input to the corresponding model to obtain the corresponding risk level.

[0077] In the aforementioned embodiment, the output of the machine learning model at each level is "risky" or "no risk". In one embodiment, Figure 2 As shown, various prediction samples are input into the machine learning models of corresponding levels in steps S510a, 511a and 512a respectively, and then in step S513a, the output results of the machine learning models at each level, namely, enterprise-level risk, position-level risk and interview-level risk are combined together as evaluation results, and the evaluation results are provided to the user in step S6a, so that the user can know whether there are risks from the three aspects of enterprise, position and interview.

[0078] In another embodiment, Figure 3 As shown, the "risky" or "no risk" output by the model is used as two levels, represented by 1 and 0 respectively, so that the risk levels of the enterprise level, position level and interview level can be encoded to obtain risk codes. After various prediction samples are input into the machine learning models of the corresponding levels to obtain the risks of the corresponding levels in steps S510a, 511a and 512a, the enterprise-level risks, position-level risks and interview-level risks are encoded in step S523a. For example, risk code 001 means that there is a risk only during the interview, while risk code 111 means that there are risks in the enterprise, position and interview. In order to enable users to feel the size of the risk, different risk levels are set in this embodiment, and the risk codes correspond to the risk levels, as shown in Table 1 below:

[0079] Table 1:

[0080]

[0081]

[0082] In step S524a, the correspondence table between risk codes and risk levels is queried based on the currently obtained risk code, as shown in Table 1. The risk level corresponding to the current risk code can be obtained and determined as the final risk level. Then, in step S6a, the final risk level is provided to the user as an evaluation result.

[0083] In another embodiment, Figure 4As shown, after encoding the obtained enterprise-level, position-level, and interview-level risk levels in step S533a, the following steps are further included:

[0084] Step S534a, obtaining the weights of the enterprise level, position level, and interview level.

[0085] Step S535a: Perform weighted calculation based on the values ​​of each digit in the risk code and their respective weights to obtain a weighted sum of the risk codes.

[0086] Step S536a, according to the current risk code weighted sum, query the corresponding table of risk code weighted sum and risk level, as shown in Table 2, to obtain the risk level that matches the current risk code weighted sum.

[0087] In one embodiment, the weights of the enterprise level, position level, and interview level are 5, 4, and 1, respectively. The weighted sums of the risk codes 000-111 are calculated to obtain 0, 1, 4, 5, 5, 6, 9, and 10. The following 8 values ​​are then divided into 3 groups based on the gaps between adjacent values, namely (0, 1), (4, 5, 6), and (9, 10), to obtain the corresponding Table 2:

[0088] Table 2:

[0089] Risk Code Weighted Sum Risk Level 9、10 High risk 4，5、6 Medium risk 0、1 Low risk

[0090] For example, in one embodiment, when the following risks are obtained according to the models of each level: enterprise level: "risky"; position level: "no risk"; interview level: "no risk", the risk code is 100, and the corresponding weighted sum of the risk codes is 5. Table 2 is queried to match the corresponding risk level of "medium risk".

[0091] Finally, in step S6a, the risk level retrieved from Table 2 is provided to the user as an evaluation result.

[0092] In the above embodiment, the model at each level adopts a binary classification output, i.e., two risk levels of "risky" and "no risk". Of course, the model at each level can also be trained to adopt a multi-classification output. For example, when five levels of "high risk", "relatively high risk", "medium risk", "relatively low risk" and "low risk" are adopted, assuming that the number of sample independent variables (i.e., the dimensions of sample features) is n, the mathematical representation of the model is:

[0093] X={x1,x2,…,x n}

[0094] Y={y0,y1,y2,y3,y4}

[0095] The X is the feature set of a sample input to the model, and each xi is a one-dimensional feature. For example, x1 is the number of employees in the enterprise, x2 is the number of branches, x3 is the registered capital, ..., x n is the annual turnover. Y is the classification set of the model, y i ∈{0,1,2,3,4}, in this example, y0 represents "low risk", y1 represents "lower risk", y2 represents "medium risk", y3 represents "higher risk", and y4 represents "high risk". The model is based on a given sample X i , calculated as label c j The probability p(c j |X i ), the output of the model is:

[0096]

[0097] in Represents the classification prediction value, p(c j |x i ) represents the probability of each classification under a given sample. In one embodiment, a comparison threshold is set, which is generally set to 0.6-0.9. In this embodiment, a threshold of 0.8 is used for illustration. For example, when judging the enterprise-level risk category, if the predicted probability of y1 output by the algorithm model exceeds 0.8, it can be confirmed that the enterprise corresponding to the sample has a lower enterprise-level risk. Among them, the threshold can be an empirical value obtained by repeatedly calculating the model based on practical data, or it can be further adjusted as the model is iterated. The trained model is verified using the validation set sample. When the model meets the evaluation criteria, it can be used for online risk assessment.

[0098] The training process of the position-level machine learning model and the interview-level machine learning model is similar to the above process and will not be repeated here.

[0099] After obtaining the risk level of each level, encoding is performed or a weighted sum of the encodings is calculated using the aforementioned embodiment to determine the final risk level, which will not be described in detail here.

[0100] Figure 5This is a flow chart for assessing risk according to another embodiment of the present invention. In this embodiment, there is one enterprise-level machine learning model, multiple position-level machine learning models and interview-level machine learning models, and the output categories of the enterprise-level machine learning model, the position-level machine learning model, and the interview-level machine learning model are more than two. The number of output categories of these three models can be the same or different. The levels are set from top to bottom in the order from enterprise level, position level to interview level. The number of next-level machine learning models is the same as the number of output categories of the previous-level machine learning model, and they are respectively trained by data with risks of the output categories of the previous-level machine learning model. For example, multiple position-level machine learning models are respectively trained by data with corresponding enterprise-level risk levels, and respectively correspond to the risk levels output by the corresponding enterprise-level machine learning models; multiple interview-level machine learning models are respectively trained by data with corresponding enterprise-level risk levels and corresponding position-level risk levels, and respectively correspond to the risk levels output by the corresponding position-level machine learning models.

[0101] For example, in one embodiment, the outputs of the enterprise-level machine learning model and the position-level machine learning model are binary classification outputs, which are defined as two risk levels of "risky" and "no risk" respectively. The interview-level machine learning model is a multi-classification output, for example, five levels of "high risk", "relatively high risk", "medium risk", "low risk" and "no risk" are defined respectively. The enterprise-level machine learning model is denoted as M1, and there are two position-level machine learning models, denoted as model M2.1 and model M2.2, respectively, wherein the position samples in the training set of model M2.1 are all samples corresponding to enterprise-level risk-free samples, and the position samples in the training set of model M2.2 are all samples corresponding to enterprise-level risky samples. There are four interview-level machine learning models, namely model M3.1, M3.2, M3.3 and model M3.4. The position samples in the training set of model M3.1 are all samples corresponding to enterprise-level risk-free and position-level risk-free samples, the position samples in the training set of model M3.2 are all samples corresponding to enterprise-level risk-free and position-level risky samples, the position samples in the training set of model M3.3 are all samples corresponding to enterprise-level risky and position-level risk-free samples, and the position samples in the training set of model M3.4 are all samples corresponding to enterprise-level risky and position-level risky samples.

[0102] The machine learning model training samples show that the enterprise-level machine learning model is the first level, the position-level machine learning model is its subordinate model, and the interview-level machine learning model is a subordinate model of the position-level machine learning model. The selection of subordinate models should correspond to the outputs of the parent model, such as risk.

[0103] In descending order of enterprise level, position level and interview level, the next level machine learning model is selected based on the risk level output by the previous level machine learning model, where the risk and level output by the interview level machine learning model are the final risk and level. The specific risk assessment process is as follows: Figure 5 As shown, the following steps are included:

[0104] Step S51a: Input the enterprise-level prediction sample into the enterprise-level machine learning model M1 for prediction. In one embodiment, the enterprise-level machine learning model is denoted as M1, and the enterprise-level prediction sample is input into the model M1.

[0105] Step S52a, determine whether the output of the enterprise-level machine learning model M1 is "risky", if it is "risky", execute step S53a, if it is "no risk", execute step S57a.

[0106] Step S53a, select the position-level machine learning model M2.2, and input the position-level prediction sample into the position-level machine learning model M2.2 for prediction.

[0107] Step S54a, determine whether the output of the position-level machine learning model M2.2 is "risky", if it is "risky", execute step S55a, if it is "no risk", execute step S56a.

[0108] Step S55a, select the interview-level machine learning model M3.4, and input the interview-level prediction sample into the interview-level machine learning model M3.4 for prediction, input the output of the interview-level prediction sample into the model M3.4 as the risk assessment result, and then end the risk assessment process.

[0109] Step S56a, select the interview-level machine learning model M3.3, and input the interview-level prediction sample into the interview-level machine learning model M3.3 for prediction, use the output of the interview-level machine learning model M3.3 as the risk assessment result, and then end the risk assessment process.

[0110] Step S57a, select the position-level machine learning model M2.1, and input the position-level prediction sample into the position-level machine learning model M2.1 for prediction.

[0111] Step S58a, determine whether the output of the position-level machine learning model M2.1 is "risky". If the output of the position-level machine learning model M2.1 is "risky", execute step S59a. If the output of the position-level machine learning model M2.1 is "no risk", execute step S510a.

[0112] Step S59a, select the interview-level machine learning model M3.2, input the interview-level prediction sample into the interview-level machine learning model M3.2 for prediction, use the output of the interview-level machine learning model M3.2 as the risk assessment result, and then end the risk assessment process.

[0113] Step S510a, select the interview-level machine learning model M3.1, input the interview-level prediction sample into the interview-level machine learning model M3.1 for prediction, use the output of the interview-level machine learning model M3.1 as the risk assessment result, and then end the risk assessment process.

[0114] This embodiment divides all information according to the overall enterprise level, position level, and interview level. The enterprise level assessment results cover the position level assessment results, and the position level assessment results cover the interview level assessment results, thereby making a comprehensive and accurate risk assessment.

[0115] In this embodiment, the interview-level machine learning model has multi-classification outputs. In one embodiment, the interview-level machine learning model obtains each classification probability by calculating the input prediction samples, for example, the probabilities of "high risk", "relatively high risk", "medium risk", "low risk" and "no risk" respectively; each classification probability is then compared with the corresponding second classification threshold; if one of the classification probabilities is greater than or equal to the corresponding second classification threshold, the model output is determined to be the risk level defined by the classification.

[0116] In another embodiment, the output of the interview-level machine learning model is binary classification; it is defined as two risk levels of "risky" and "no risk" respectively; when calculating the predicted sample, the interview-level machine learning model calculates the "risky" probability; and compares the "risky" probability with multiple third classification thresholds, wherein the multiple third classification thresholds constitute multiple risk probability intervals from low to high, corresponding to multiple risk levels from low to high respectively; the corresponding risk level is determined according to the risk probability interval in which the "risky" probability output by the interview-level machine learning model is located.

[0117] In another embodiment, the outputs of the enterprise-level machine learning model and the position-level machine learning model can also be multi-classified outputs, similar to the outputs of the interview-level machine learning model. Each position-level machine learning model corresponds to a class of outputs of the enterprise-level machine learning model, and thus the number of position-level machine learning models is the same as the number of risk output classes of the enterprise-level machine learning model. Correspondingly, the number of interview-level machine learning models is the product of the number of position-level machine learning models and the number of risk output classes of the position-level machine learning models.

[0118] In step S6a, the determined risk is provided to the user as an identification result, and the providing manner includes displaying it in a terminal interface, sending an email, sending a short message through a mobile communication network, or sending a social message to the user's social media account.

[0119] Figure 6 FIG. 1 is a flow chart of a method for processing prediction sample data according to an embodiment of the present invention. In this embodiment, the method includes the following steps:

[0120] Step S1b monitors user feedback on each recognition result. After evaluating the job interview information provided by the user to determine whether the current job opening is a fraudulent one and providing the recognition result to the user, user feedback on the recognition result is monitored. In one embodiment, the feedback may include confirmation of whether the recognition result is "risky" or "risk-free," or evaluation of whether the recognition result is "correct" or "incorrect."

[0121] Step S2b: Extract the user's confirmation information on the recognition result. Based on the information fed back in step S1b, extract the confirmation information of "risky" or "no risk", or the evaluation information of "correct" or "wrong".

[0122] Step S3b, sets corresponding labels for corresponding prediction samples based on the user's confirmation information. For example, when the user confirms that the job interview input by the user is a real job, the labels of real job are set for the prediction samples at all levels that are evaluated for the job. If the user confirms that the job interview input by the user is a false job, the labels of false job are set for the prediction samples at all levels that are evaluated for the job. In a better embodiment, the user's feedback information also includes more information fields. For example, when the user confirms that it is a false job, the user needs to fill in the reason. The present invention analyzes the reasons filled in by the user to evaluate the root cause of the falsehood from the two aspects of "enterprise" and "position", and sets labels for the current prediction samples at all levels.

[0123] Step S4b: storing the labeled prediction samples into the training set, that is, the system stores the labeled prediction samples into the data set used for training the model, thereby enriching the training data.

[0124] Step S5b determines whether a model update condition has been met. The update condition may be, for example, a preset update cycle, such as weekly or monthly model updates, or counting the number of newly added training samples in the model training dataset and determining whether the number of newly added training samples has reached a threshold. If the update cycle has been reached or the number of newly added training samples has reached the threshold, the model update condition is determined to be met, and step S6b is executed. Otherwise, the process returns to step S1b.

[0125] Step S6b: Use the training data set to optimize and update the currently used machine learning model.

[0126] The present invention can continuously accumulate training data without the need to manually set labels for the training data. The model optimization can be carried out automatically without the need for human intervention, thus saving a lot of manpower and time. Moreover, as the model is optimized, the accuracy of the model evaluation is gradually improved, and false recruitment can be identified more accurately, thereby better safeguarding the interests of users.

[0127] In one embodiment, the present invention also provides a false recruitment warning method, see Figure 7 , Figure 7 The flowchart of the false recruitment warning method according to one embodiment of the present invention includes:

[0128] Step S1c, determining whether the recruitment interview information provided by the user is received, if the recruitment interview information provided by the user is received, executing step S2c, if not, repeating this step.

[0129] Step S2c: extracting target recruiting company information, target position information and interview information from the recruitment interview information provided by the user.

[0130] Step S3c: crawling information on the network. Based on the extracted target recruiting company information, target position information, and interview information, crawling information on the network is performed according to multiple indicators of company level, position level, and interview level.

[0131] Step S4c: Process the crawled information into indicator data of the corresponding level.

[0132] Step S5c, perform risk assessment based on multiple indicator data at each level. Figure 2-5 Perform a risk assessment using any of the processes shown.

[0133] Step S6c: Based on the assessed risk, trace back and analyze the indicator data corresponding to the risk to obtain warning information and provide it to the user. The warning information includes the problem indicator data causing the risk and the risk content.

[0134] The steps for reviewing and analyzing indicator data corresponding to the risk include:

[0135] First, according to the monitored risk type, the indicator data used in assessing the risk are traversed to calculate the contribution of each indicator data to the assessed risk.

[0136] Then, the multiple indicators are sorted according to their contribution to the risk, and the multiple indicators with the highest sorting or indicators with a contribution greater than a threshold are determined as abnormal indicators.

[0137] Finally, risk content is generated based on the content of the abnormal indicator and / or the association relationship between the contents of multiple abnormal indicators.

[0138] Among them, the feature importance determination method in feature engineering can be used to determine the contribution of indicator data to the assessed risk. The feature importance determination method adopts, for example, the expert meeting method, in which experts in this field specify and identify the importance of each indicator used in the present invention, or adopts rough set theory, information entropy and other methods to identify the importance of each indicator used in the present invention, or uses data mining technology to mine the correlation between indicators and outputs to determine the importance of indicators. Based on the indicator importance determined by the above methods, when backtracking the indicators, the contribution is determined by querying the importance identification corresponding to each indicator.

[0139] There are other methods for determining feature importance that differ from the above. For example, when using a neural network machine learning model, since the weights of the hidden layer nodes in the neural network represent the importance of the features of the corresponding input layer nodes, when the machine learning model uses a neural network model, the weights of the hidden layer nodes can be read and the contribution of the indicator can be determined based on the weights of the hidden layer nodes.

[0140] For another example, for a machine learning model that uses a decision tree or random forest algorithm, according to the decision tree generation principle, the order of features selected in the process of dividing the decision tree can be used as the importance ranking of the features, and the ranking can be obtained through the feature_importances_ attribute in sklearn. Therefore, when the present invention traces back indicators, the importance of each indicator used in risk assessment, that is, the contribution, can be obtained by calling the feature_importances_ attribute.

[0141] After determining the contribution of the indicators, the indicators with the highest contribution ranking (such as the top three indicators) can be identified as abnormal indicators. After identifying the abnormal indicators, the content of the abnormal indicators is read to locate the points where risks may arise and the specific risk content. The content of the three abnormal indicators is analyzed to see if there is a correlation. If so, the correlated content is also identified as risk content.

[0142] The present invention presets the risk events and their corresponding response plans corresponding to various risk contents. Therefore, after the risk content is determined, the database is searched according to the risk content to determine the corresponding response plan.

[0143] Taking the assessment of corporate risk as an example, if it is found after retrospective analysis that "corporate legal proceedings" are the main cause of corporate risk, the corresponding solutions provided by the system include: (1) During the interview, you can ask HR for verification on the company's labor dispute litigation; (2) Before the interview or joining the company, please collect more relevant information online to understand the details of the relevant litigation and strengthen your comprehensive understanding of this issue, etc.

[0144] The risk content and corresponding solutions are added to the warning information and presented to the user. Furthermore, to help the user better understand the identified content or facilitate future inquiries, the present invention can also generate a risk report based on the content in the warning information and store it or provide it to the user. For example, a risk report based on the assessed risk level, risk content, and corresponding solutions can be generated and sent to the user or stored in the user's account.

[0145] Figure 8 The following is a block diagram of a system for identifying fake job postings according to one embodiment of the present invention. In this embodiment, the system comprises a user information acquisition module 1, a data collection module 2, an indicator data generation module 3, and a risk assessment module 4. The user information acquisition module 1 is connected to the data collection module 2. The user information acquisition module 1 receives user-provided job interview information and extracts target company information, target position information, and interview information from the user's job interview information. The data collection module 2 is connected to the internet and to the user information acquisition module 1. Based on the extracted target company information, target position information, and interview information, the data collection module 2 crawls the network according to multiple indicators at the company level, position level, and interview level to obtain relevant information such as the recruiting company and the target position. The indicator data generation module 3 is connected to the data collection module 2 and processes the crawled company-level, position-level, and interview-level information to obtain multiple indicator data at the corresponding levels. After processing by the indicator data generation module 3, the information at each level is normalized into specific indicator data, such as floating-point numbers, string vectors, or some encoding. The risk assessment module 4 is connected to the indicator data generation module 3 and is configured to evaluate the target enterprise's false recruitment risk, position-level risk and interview-level risk based on multiple indicator data at various levels.

[0146] Figure 9 This is a partial principle block diagram of a fake recruitment identification system according to another embodiment of the present invention. This embodiment uses a machine learning model to assess risk. Figure 8In addition to the modules in the example, the system also includes a prediction sample generation module 5, which is connected to the indicator data generation module 3. Based on the samples required by the machine learning models at each level, each indicator data is normalized into feature data, and the feature data corresponding to the indicator data at each level are combined to form prediction samples of the machine learning models at each level. The risk assessment module 4 in this embodiment includes an enterprise-level risk assessment unit 41, a position-level risk assessment unit 42, an interview-level risk assessment unit 43, a coding unit 44, and a risk level query unit 45. The prediction sample generation module 5 inputs the prediction samples at each level into the enterprise-level risk assessment unit 41, the position-level risk assessment unit 42, and the interview-level risk assessment unit 43 respectively. The enterprise-level risk assessment unit 41 obtains the enterprise-level risk probability by inputting the enterprise-level prediction samples into the trained enterprise-level machine learning model. The position-level risk assessment unit 42 obtains the position-level risk probability by inputting the position-level prediction samples into the trained position-level machine learning model. The interview-level risk assessment unit 43 obtains the interview-level risk probability by inputting the interview-level prediction samples into the trained interview-level machine learning model. In one embodiment, the risk probabilities obtained by these three units can be directly provided to the user as identification results. In this embodiment, the three units output the obtained risk probabilities to the encoding unit 44, which encodes the obtained enterprise-level, position-level, and interview-level risk levels to obtain risk codes, and outputs the risk codes to the risk level query unit 45. The risk level query unit 45 queries the corresponding table between codes and risk levels based on the risk codes, such as Table 1 involved in the aforementioned description of the method, and determines the risk level corresponding to the risk code as the final risk level. Of course, after obtaining the risk code, the encoding unit 44 can also calculate the weighted sum of the risk codes based on the weights of the risks at each level. The risk level query unit 45 queries the corresponding table between the weighted sum of the codes and risk levels based on the weighted sum of the risk codes, such as Table 2 involved in the aforementioned description of the method, and determines the risk level corresponding to the weighted sum of the risk codes as the final risk level.

[0147] Figure 10 This is a block diagram of the principle of a fake recruitment identification system according to another embodiment of the present invention. Figure 9Compared with the embodiment shown in the figure, the risk assessment module 4 in this embodiment includes not only an enterprise-level risk assessment unit 41, a position-level risk assessment unit 42 and an interview-level risk assessment unit 43, but also a selection unit 46. In this embodiment, the machine learning models used by the three assessment units are related to each other. Among them, in the order of enterprise level, position level and interview level from top to bottom, the next-level machine learning model is trained according to the training data corresponding to the risk level output by its upper-level machine learning model. Therefore, the number of the next-level machine learning models is the same as the number of risk levels of the upper-level machine learning model, and the risk assessment is also evaluated step by step from top to bottom, and the next-level machine learning model needs to be selected according to the risk level of the upper-level machine learning model.

[0148] Specifically, the enterprise-level risk assessment unit 41 receives the enterprise-level prediction samples obtained by the prediction sample generation module 5 and notifies the selection unit 46. The selection unit 46 selects an enterprise-level machine learning model from the model library and sends it to the enterprise-level risk assessment unit 41. The enterprise-level risk assessment unit 41 inputs the enterprise-level prediction samples into the enterprise-level machine learning model, obtains the enterprise-level risk level through evaluation by the enterprise-level machine learning model, sends the enterprise-level risk level to the selection unit 46, and notifies the position-level risk assessment unit 42. The selection unit 46 selects an appropriate position-level machine learning model based on the enterprise-level risk level and sends it to the position-level risk assessment unit 42. After receiving the notification from the enterprise-level risk assessment unit 41 and the position-level machine learning model sent by the selection unit 46, the position-level risk assessment unit 42 inputs the position-level prediction samples received from the prediction sample generation module 5 into the position-level machine learning model, obtains the position-level risk level through evaluation by the position-level machine learning model, outputs the position-level risk level to the selection unit 46, and simultaneously outputs a notification to the interview-level risk assessment unit 43. The selection unit 46 selects the corresponding interview-level machine learning model according to the position-level risk level and sends it to the interview-level risk assessment unit 43. After receiving the notification from the position-level risk assessment unit 42 and the interview-level machine learning model sent by the selection unit 46, the interview-level risk assessment unit 43 inputs the interview-level prediction sample received from the prediction sample generation module 5 into the interview-level machine learning model, and the final level risk level is obtained through evaluation by the interview-level machine learning model.

[0149] The outputs of the enterprise-level machine learning model and the position-level machine learning model in the aforementioned embodiment are each binary, defined as two risk levels: "risky" and "no risk." The output of the interview-level machine learning model is either binary or multi-classified. When binary, the outputs are defined as two risk levels: "risky" and "no risk." When multi-classified, the outputs are defined as multiple risk levels ranging from zero to risky.

[0150] Figure 11 The figure is a functional block diagram of a user information acquisition module according to one embodiment of the present invention. In this embodiment, the user information acquisition module includes a message transceiver unit 11 and an information extraction unit 12. The message transceiver unit 11 serves as the interface between the system and the user. On the one hand, it connects to the risk assessment module 4 to send the final recognition results to the user. On the other hand, it receives recruitment interview information provided by the user and outputs it to the information extraction unit 12. The information extraction unit 12 extracts target company information, target position information, and interview information from the recruitment interview information provided by the user.

[0151] The message transceiver unit 11 includes one or more of the following units: an application terminal user interaction unit 110, an email processing unit 111, a mobile short message processing unit 112, and a social media message processing unit 113. The application terminal user interaction unit 110 includes at least an input interface, through which recruitment interview information provided by users via the application terminal can be obtained. Additionally, the application terminal user interaction unit 110 may also include a display interface for displaying messages, such as identified risk levels, risk content, or corresponding solutions. The email processing unit 111 can identify recruitment interview information provided by users via email based on the email address or subject line, and can also send messages to users via email. The mobile short message processing unit 112 can identify recruitment interview information provided by users via short messages based on the message sender number and subject line from short messages received over the mobile communication network, and can also send messages to users. The social media message processing unit 113 can identify recruitment interview information provided by users via social media based on the message sender, or send messages to users.

[0152] Figure 12 The data processing system in this embodiment includes a user information acquisition module 1, a data collection module 2, an indicator data generation module 3 and a prediction sample generation module 5. Figure 9 or Figure 10 Some modules in the fake recruitment identification system constitute a data processing system, which performs information extraction, information crawling, data range analysis and other processing based on the recruitment interview information provided by users, thereby obtaining prediction samples for machine learning models to assess risks. Specifically, Figure 13 FIG. 2 is a block diagram of a data collection module according to an embodiment of the present invention. In this embodiment, the data collection module includes an indicator acquisition unit 21 , an indicator analysis unit 22 , an information crawling unit 23 and an information matrix construction unit 24 .

[0153] The indicator acquisition unit 21 is used to read multiple indicators that are applied to the enterprise level, position level, and interview level respectively. In this embodiment, the system stores indicators for assessing risks at various levels. The indicator acquisition unit 21 reads these indicators from the system database and sends them to the indicator analysis unit 22. The indicator analysis unit 22 analyzes each indicator, determines the indicator reference content required to obtain the indicator data, and sends the indicator reference content to the information crawling unit 23. The information crawling unit 23 is connected to the indicator analysis unit 22 and crawls corresponding information from the Internet based on the determined indicator reference content. In one embodiment, the system stores one or more search keywords corresponding to the indicators. For example, when the indicator acquisition unit 21 reads the indicator "enterprise headcount", the indicator analysis unit 22 queries its search keyword to obtain the indicator reference content "enterprise headcount / quantity". The information crawling unit 23 searches the target recruiting company's official website based on the indicator reference content "enterprise headcount / quantity" and obtains the company headcount information to obtain information that meets the indicator "enterprise headcount". For another example, when the indicator acquisition unit 21 reads the enterprise-level indicator "enterprise legal litigation events", the indicator analysis unit 22 queries its search keywords to obtain the indicator reference content "labor / defendant", etc., and the information crawling unit 23 queries the target recruiting enterprise for relevant litigation information in the recent period (such as five years) based on these contents.

[0154] In this embodiment, the information matrix construction unit 24 is connected to the information crawling unit 23 and creates an information matrix based on the relationships between the crawled information. The information matrix includes one or more job postings published by the target enterprise and the qualification information of the recruitment platforms that publish these job postings; and job postings published by one or more similar enterprises to the target enterprise and the qualification information of the recruitment platforms that publish these job postings. The job postings published by the target enterprise include the target position and a second position different from the target position; and the recruitment platforms include one or more target recruitment platforms that publish the target position provided by the target enterprise and one or more second recruitment platforms that publish the second position information different from the target position.

[0155] Figure 14A block diagram of the principle of the indicator data generation module according to one embodiment of the present invention. In this embodiment, the indicator data generation module includes a data cleaning unit 31, a single indicator extraction unit 32, and a composite indicator calculation unit 33. The data cleaning unit 31 performs data cleaning on the crawled raw information, including removing certain network symbols and punctuation marks, and querying a stop word list to remove stop words. The single indicator extraction unit 32 is connected to the data cleaning unit 31 to extract and normalize single indicator data from the cleaned data. Examples of single indicators include those that do not require complex calculations or processing, such as "enterprise headcount," "number of branches," "registered capital," "position title," and "educational requirements." For these indicators, indicator data can be extracted by identifying keywords and, depending on the specific indicator, processed into string vectors, floating-point values, or codes. The composite indicator calculation unit 33 is connected to the single indicator extraction unit 32 and configured to calculate one or more single indicator data according to composite indicator calculation rules to generate composite indicator data. Examples of composite indicators include "recruitment platform commitment," "recruitment platform qualification level," and "external consistency coefficient of target position." The calculation method is as described in the method section above and will not be repeated here.

[0156] Figure 15: This is a principle block diagram of a model training module in a data processing system provided according to another embodiment of the present invention. In this embodiment, the model training module 6 includes a training data set unit 61, a model training unit 62, a user feedback monitoring unit 63, a sample labeling unit 64, and a model updating unit 65. The training data set unit 61 is used to provide a data set for training the model, and includes an enterprise-level training data subset, a position-level training data subset, and an interview-level training data subset according to the type of model to be trained. Furthermore, each training data subset also includes a training subset and a validation subset. The model training unit 62 performs model training with the data in the corresponding type of training data subset according to the type of model to be trained, and obtains an enterprise-level machine learning model, a position-level machine learning model, and an interview-level machine learning model respectively. In one embodiment, the model training unit 62 performs model training with independent training sets to obtain three independent machine learning models. In another embodiment, according to the risk level of the enterprise-level machine learning model, the position-level training data subsets respectively include different subsets consisting of training data with corresponding enterprise risks, and different position-level machine learning models are obtained according to different training subsets. Similarly, corresponding to the interview-level training data subset, which includes multiple training subsets consisting of training data specifically corresponding to the enterprise risk level and the position-level risk level, different interview-level machine learning models are obtained according to different training subsets. As mentioned above, when explaining the method of the present invention, the enterprise-level machine learning model M1, the position-level machine learning models M2.1, M2.2, and the interview-level machine learning models M3.1, M3.2, M3.3 and M3.4 are obtained through different training subsets. In order to expand the training set, the user feedback monitoring unit 63 monitors the user's feedback information on the false recruitment identification results, and the feedback information includes at least confirmation information of risk or no risk. The sample annotation unit 64 is connected to the user feedback monitoring unit 63, and based on the user's feedback information, the data for obtaining the identification results is risk-labeled and added to the corresponding training data subset, thereby achieving the purpose of enriching the training data. The model update unit 65 is connected to the training data set unit 61, and is used to monitor the update conditions. When the model update conditions are met, an update notification is sent to the model training unit 62. The model training unit 62 uses the training data to train, optimize, and update the model. The update conditions include, for example, reaching a preset update cycle, such as updating the model weekly / monthly, or reaching a threshold for new training samples. Therefore, the model update unit 65 counts after each model update and sends an update notification to the model training unit 62 when the timing cycle is reached. Alternatively, the model update unit 65 counts the newly added training data in the training dataset and, when the amount of newly added training data reaches a threshold, notifies the model training unit to optimize and update the original machine learning model using the current training data.

[0157] Figure 16 The following is a block diagram of the principles of a fraudulent recruitment early warning system according to one embodiment of the present invention. In this embodiment, the early warning system includes a user information acquisition module 1, a data collection module 2, an indicator data generation module 3, a risk assessment module 4, and an early warning module 7. The user information acquisition module 1, data collection module 2, indicator data generation module 3, and risk assessment module 4 are identical to those in the fraudulent recruitment identification system described in the previous embodiment and are not further described here.

[0158] The early warning module 7 is connected to the risk assessment module 4. In response to the assessed risk, the early warning module 7 back-traces and analyzes the indicator data corresponding to the risk to obtain early warning information and provide it to the user.

[0159] Figure 17 7 is a block diagram of the principle of an early warning module according to an embodiment of the present invention. In this embodiment, the early warning module includes a risk monitoring unit 71, a risk content determination unit 72, an early warning information generation unit 73 and an early warning information sending unit 74.

[0160] The risk monitoring unit 71 is connected to the risk assessment module 4 to monitor whether the risk assessment module 4 has assessed a risk. When it is detected that the risk assessment module 4 has assessed a risk, a notification is sent to the risk content determination unit 72. The risk content determination unit 72 traverses the indicator data used to assess the risk based on the assessed risk type to obtain abnormal indicator data, and determines the risk content based on the content of the one or more abnormal indicator data or their correlation. Among them, any one of the aforementioned methods can be used when determining the abnormal indicator data, which will not be repeated here. The warning information generation unit 73 is connected to the risk content determination unit 72 to generate warning information based on the risk type and risk content. The warning information sending unit 74 is connected to the warning information generation unit 73 to provide the warning information to the user. Among them, the warning information sending unit 74 can adopt one or more of the interactive information sending unit, email processing unit, mobile short message processing unit, mobile short message processing unit and social media message processing unit. That is to say, the warning information sending unit 74 can be merged with the user information acquisition module into one module, so that its module is implemented as a two-way function with both message receiving and sending. Please refer to the above embodiment for details and will not be repeated here.

[0161] Figure 18 This is a block diagram of the principle of the early warning module according to another embodiment of the present invention. Figure 17Compared with the embodiment, except for the addition of the prediction unit 75, the functions of other modules are the same and will not be repeated here. The prediction unit 75 is connected to the risk content determination unit 72, and is configured to predict risk events and risk response plans based on the risk content, and add the risk events and their response plans to the early warning information. For example, the prediction unit 75 queries the various suggestions and corresponding plans pre-set in the system database based on the risk content, thereby obtaining matching suggestions, and combines multiple suggestions or corresponding plans together and adds them to the early warning information. For another example, the system database is pre-set with various possible risk events and corresponding response plans corresponding to the risk content. This information can be added to the early warning information and sent to the user together.

[0162] Figure 19 This is a block diagram of the principle of the early warning module according to another embodiment of the present invention. Figure 18 Compared to the embodiment of FIG1 , except for the addition of a risk report generation unit 76, the functions of the other modules remain the same and are not further described here. The risk report generation unit 76 is connected to the warning information generation unit 73 and is configured to generate a risk report based on the content of the warning information. Correspondingly, the warning information transmission unit stores the risk report in a predetermined location or transmits it to the user.

[0163] The above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the scope of the present invention. Therefore, all equivalent technical solutions should also fall within the scope of the present invention.

Claims

1. A data processing method for identifying false recruitment, comprising: Extract target recruiting company information, target position information and interview information from the recruitment interview information provided by the user; Based on the extracted target recruitment company information, target position information and interview information, information is crawled on the network according to multiple indicator reference contents at the company level, position level and interview level; Based on the crawled information at each level, multiple indicator data of the corresponding level are obtained; and Normalize the indicator data into feature data of the machine learning model, and combine the feature data corresponding to the indicator data at all levels to form the prediction samples of the machine learning model at all levels; Among them, the position-level indicator data includes depth indicator data and breadth indicator data, and an enterprise network recruitment information matrix is ​​established according to the position-level information; the breadth indicator data includes position-level breadth indicator data obtained based on the enterprise network recruitment information matrix, and the breadth indicator data includes the qualification level vector weight of the recruitment platform, the external consistency coefficient of the target position, the internal consistency coefficient of the target position, and the consistency coefficient of the target position; The enterprise-level prediction samples are input into the trained enterprise-level machine learning model to output whether the target recruiting enterprise is an enterprise with enterprise-level risk or a enterprise without enterprise-level risk. When the target recruiting company is a risk-free company at the enterprise level, the position-level prediction samples are input into the trained position-level machine learning model M2.1 to output whether the target recruiting company is a risky company at the position level or a risk-free company at the position level; When the target recruiting company is a company with enterprise-level risk, the position-level prediction sample is input into the trained position-level machine learning model M2.2 to output whether the target recruiting company is a company with position-level risk or a company with no position-level risk. When the target recruiting company is risk-free at both the company level and the position level, the interview-level prediction samples are input into the trained interview-level machine learning model M3.1 to output whether the target recruiting company is a risky company at the interview level or a risk-free company at the interview level. When the target recruiting company is risk-free at the enterprise level and risky at the position level, the interview-level prediction samples are input into the trained interview-level machine learning model M3.2 to output whether the target recruiting company is a risky company at the interview level or a risk-free company at the interview level. When the target recruiting company is a company with enterprise-level risk but no risk at the position level, the interview-level prediction sample is input into the trained interview-level machine learning model M3.3 to output whether the target recruiting company is a company with interview-level risk or a company with no risk at the interview level. When the target recruiting enterprise is an enterprise with risks at the enterprise level and job level, the interview-level prediction sample is input into the trained interview-level machine learning model M3.4 to output whether the target recruiting enterprise is an enterprise with risks at the interview level or the target recruiting enterprise is an enterprise without risks at the interview level.

2. The method according to claim 1, wherein the enterprise-level indicator data obtained from the crawled enterprise-level information includes one or more static indicator data of the target recruiting enterprise, including the number of employees, number of branches, registered capital, financing information, company years, annual turnover, business scope and company type.

3. The method according to claim 2, wherein the enterprise-level indicator data obtained from the crawled enterprise-level information also includes one or more dynamic indicator data of corporate legal proceedings, major investments received, social news events, and social media corporate evaluations of the target recruiting enterprise.

4. The method according to claim 3, wherein the dynamic indicator data obtained further comprises: The relevance of each dynamic information to recruitment is calculated, and the weight of the dynamic indicator data is set according to the relevance.

5. The method according to claim 1, wherein the depth indicator data comprises: One or more of the target position's job title, job location, job department, monthly salary range, job description, benefits, job type, experience requirements, educational requirements, school type, job industry risk, and the number of repeated recruitments for the same position.

6. The method according to claim 5, wherein the enterprise network recruitment information matrix comprises: One or more job postings posted by the target enterprise and qualification information of the recruitment platform that posted the job postings, the job postings including the target job and a second job posting different from the target job; Job information published by one or more similar enterprises to the target enterprise and qualification information of the recruitment platforms that publish the job information, wherein the recruitment platforms include the target recruitment platform and the second recruitment platform that publish the target job provided by the target enterprise.

7. The method according to claim 6, wherein the step of obtaining the weighted vector value of the qualification level of the recruitment platform based on the enterprise network recruitment information matrix comprises: Calculating a qualification level vector for each recruitment platform based on the recruitment platform qualification information; Count the number of positions posted on each recruitment platform from the enterprise network recruitment information matrix; Determine the level of commitment to each recruitment platform based on the number of positions stated; as well as The degree of investment in each recruitment platform is used as the qualification level vector weight of the recruitment platform, and the qualification level weighted vector value of each recruitment platform is calculated.

8. The method according to claim 7, wherein the step of obtaining the external consistency coefficient of the target position based on the enterprise network recruitment information matrix comprises: Searching for multiple second target positions that are the same as the target position and are provided by similar companies; Calculate a position vector of the target position and each second target position based on the position information; performing a clustering operation on the plurality of second target positions to obtain a second target position with the largest clustering result; as well as Calculate a first vector difference between the position vector of the target position and the position vector of the second target position with the largest clustering result, and use the first vector difference as the external consistency coefficient of the target position.

9. The method according to claim 8, wherein the step of obtaining the internal consistency coefficient of the target position based on the enterprise network recruitment information matrix comprises: Calculate the second vector difference between the position vector of the target position and the position vector of each second position provided by the target enterprise, and calculate the average value of the second vector differences, and use the average value of the second vector differences as the internal consistency coefficient of the target position.

10. The method according to claim 9, wherein the step of obtaining the consistency coefficient of the target position based on the enterprise network recruitment information matrix comprises: Based on the average value of the first vector difference and the second vector difference, the respective weights are determined according to their respective closeness relationships with the target position vector, and the weighted average value of the two is calculated as the consistency coefficient of the target position.

11. The method according to claim 1, wherein The steps to obtain interview-level indicators based on the crawled interview-level information include: Obtain the geographic location information and owner / tenant information of the building where the interview location is located based on the interview location; Matching the building's owner / tenant information with the target business name; and Determine interview location indicators based on matching results.

12. The method according to claim 11, wherein The interview level indicators also include one or more of interview time, interview notification method, number of interviews and interview format.

13. The method of claim 1 , further comprising: Monitoring user feedback on fake recruitment identification results, wherein the feedback information at least includes confirmation information of risk or no risk; as well as According to the risky or non-risky confirmation information in the feedback information, corresponding labels are set for the prediction samples provided to the machine learning models at all levels and stored in the model training data set.

14. The method according to claim 13, further comprising: Monitor whether the amount of new training data reaches the threshold; as well as In response to the amount of newly added training data reaching a threshold, the currently used machine learning model is optimized and updated.

15. A data processing system for identifying false recruitment, comprising: A user information acquisition module is configured to receive recruitment interview information provided by a user and extract target recruiting company information, target position information and interview information from the user's recruitment interview information; a data collection module connected to the Internet and to the user information acquisition module, configured to crawl information from the network based on the extracted target recruiting company information, target position information, and interview information, according to multiple indicator reference contents at the company level, position level, and interview level; An indicator data generation module, which is connected to the data collection module and processes the crawled information at each level to obtain multiple indicator data at corresponding levels; and a prediction sample generation module, connected to the indicator data generation module, configured to normalize the indicator data into feature data of the machine learning model, and combine the feature data corresponding to the indicator data at each level to form prediction samples for the machine learning model at each level; Among them, the position-level indicator data includes depth indicator data and breadth indicator data, and an enterprise network recruitment information matrix is ​​established according to the position-level information; the breadth indicator data includes position-level breadth indicator data obtained based on the enterprise network recruitment information matrix, and the breadth indicator data includes the qualification level vector weight of the recruitment platform, the external consistency coefficient of the target position, the internal consistency coefficient of the target position, and the consistency coefficient of the target position; The enterprise-level prediction samples are input into the trained enterprise-level machine learning model to output whether the target recruiting enterprise is an enterprise with enterprise-level risk or a enterprise without enterprise-level risk. When the target recruiting company is a risk-free company at the enterprise level, the position-level prediction samples are input into the trained position-level machine learning model M2.1 to output whether the target recruiting company is a risky company at the position level or a risk-free company at the position level; When the target recruiting company is a company with enterprise-level risk, the position-level prediction sample is input into the trained position-level machine learning model M2.2 to output whether the target recruiting company is a company with position-level risk or a company with no position-level risk. When the target recruiting company is risk-free at both the company level and the position level, the interview-level prediction samples are input into the trained interview-level machine learning model M3.1 to output whether the target recruiting company is a risky company at the interview level or a risk-free company at the interview level. When the target recruiting company is risk-free at the enterprise level and risky at the position level, the interview-level prediction samples are input into the trained interview-level machine learning model M3.2 to output whether the target recruiting company is a risky company at the interview level or a risk-free company at the interview level. When the target recruiting company is a company with enterprise-level risk but no risk at the position level, the interview-level prediction sample is input into the trained interview-level machine learning model M3.3 to output whether the target recruiting company is a company with interview-level risk or a company with no risk at the interview level. When the target recruiting enterprise is an enterprise with risks at the enterprise level and job level, the interview-level prediction sample is input into the trained interview-level machine learning model M3.4 to output whether the target recruiting enterprise is an enterprise with risks at the interview level or the target recruiting enterprise is an enterprise without risks at the interview level.

16. The system according to claim 15, wherein: The user information acquisition module includes: A message receiving unit configured to receive recruitment interview information provided by a user; and The information extraction unit is connected to the message receiving unit and is configured to extract target recruiting company information, target position information and interview information from the recruitment interview information provided by the user.

17. The system according to claim 16, wherein: The message receiving unit is one or more of the following units: An application terminal user interaction unit, which includes at least an input interface and is configured to obtain recruitment interview information provided by a user via the application terminal through the input interface; An email processing unit configured to identify recruitment interview information provided by a user via email based on an email address and / or subject; A mobile short message processing unit configured to identify recruitment interview information provided by a user via a short message according to a message sending number and a subject in a short message received over a mobile communication network; as well as The social media message processing unit is configured to identify recruitment interview information provided by a user via social media based on a message sender from the social media message.

18. The system according to claim 15, wherein: The data collection module includes: an indicator acquisition unit configured to read a plurality of indicators respectively applied to the enterprise level, the position level and the interview level; an indicator analysis unit connected to the indicator acquisition unit, configured to analyze each indicator and determine the indicator reference content required to obtain the indicator data; and An information crawling unit is connected to the indicator analysis unit and is configured to crawl corresponding information from the Internet according to the determined indicator reference content.

19. The system of claim 18, wherein the data collection module further comprises: The information matrix construction unit is connected to the information crawling unit and is configured to build an information matrix according to the association between the crawled information.

20. The system of claim 15, wherein: The indicator data generation module includes: A data cleaning unit, configured to perform data cleaning on the crawled raw information; a single indicator extraction unit, connected to the data cleaning unit, configured to extract and normalize single indicator data from the cleaned data; and The composite indicator calculation unit is connected to the single indicator extraction unit and is configured to calculate more than one single indicator data according to the composite indicator calculation rule to obtain composite indicator data.

21. The system according to claim 15, further comprising a model training module configured to include: A training data set unit is configured to provide a data set for training a model, including an enterprise-level training data subset, a position-level training data subset, and an interview-level training data subset according to the type of model being trained; as well as The model training unit is configured to perform model training based on the model type to be trained using data in the corresponding type of training data subset to obtain an enterprise-level machine learning model, a position-level machine learning model, and an interview-level machine learning model respectively.

22. The system of claim 21, wherein: In the training data set unit, the order is enterprise level, position level, and interview level from top to bottom, the number of each next-level subset is the same as the number of risk levels of the previous-level machine learning model, and each next-level subset is composed of training data with risks of the corresponding risk level of the previous level; or, the enterprise-level training data subset, the position-level training data subset, and the interview-level training data subset are independent data sets.

23. The system of claim 22, wherein the model training module further comprises: A user feedback monitoring unit configured to monitor user feedback on a false recruitment identification result, wherein the feedback information includes at least a confirmation information of risk or no risk; as well as A sample labeling unit is connected to the user feedback monitoring unit and is configured to label the risk of the predicted samples of the recognition result based on the user's feedback information, and add the labeled predicted samples to the corresponding training data set.

24. The system of claim 23, wherein the model training module further comprises: A model updating unit is connected to the training data set unit and is configured to count the new training data in the training data set. When the amount of new training data reaches a threshold, the model training unit is notified to optimize and update the original machine learning model with the current training data.

Images